Crop Disease Diagnosis Model Using Deep Hybrid Architecture With a New Segnet‐Based Segmentation Model

The productivity of common crops in agriculture is declining, leading to increased interest in alternative crops known for their adaptability, nutritional value and resilience to harsh climates. Despite their benefits, these crops face threats from diseases that harm the agricultural economy. Detecting the disease using traditional methods is labor‐intensive, costly, and often questioned for their accuracy. To overcome these challenges, a novel Crop Disease Diagnosis framework using an Enhanced Convolutional Neural Network (CDD‐EC) model has been proposed. In data acquisition, the images from a benchmark dataset are collected using the range of plant disease samples in the initial stage of the CDD‐EC procedure. Gaussian filtering is used as a preprocessing technique on these photos to reduce noise and enhance image quality. A Coordinate Attention module‐assisted SegNet (CASNet) model is then used for segmentation, effectively separating the sick areas from the background. Following segmentation, the Pyramid Histogram of Oriented Gradients (PHOG) is used to analyse the structure and spatial distribution, extract texture features using the Updated Gabor Filter in Local Gabor Directional Pattern (UGF‐LGDP), and derive structural shape information. In a Bidirectional Long Short‐Term Memory network (Bi‐LSTM), the combined features are processed to learn both backward and forward dependencies in the data. The Normalized Average pooling response Batch normalization‐based Convolutional Neural Network (NAB‐CNN) processes the output feature map of Bi‐LSTM for additional refinement and classification. Finally, the system produces a categorized output that identifies the specific type of plant disease present in the image. The CDD‐EC scheme attained the highest accuracy of 0.937, precision of 0.875 and specificity of 0.958. Also, it demonstrated reduced error values with FDR at 0.125 and FPR at 0.042, respectively.